Psyche


Psyche is a decentralized network for coordinating distributed model training across heterogeneous compute.

How it works

Psyche shards optimizer state and gradient computation across volunteer and datacenter nodes alike, using a fault-tolerant coordination protocol that tolerates churn, latency, and partial failure.

Why it matters

Training frontier models should not require a single hyperscale datacenter. Psyche makes large-scale, unbiased pretraining a public good.